The Inference Report

The SWE-rebench rankings hold steady at the top, with Anthropic's Fable 5 maintaining 64.5% and the same seven models occupying positions one through seven as before. Grok 4.5 remains second at 63.8%, Opus 5 third at 63.4%, and the ordering unchanged through Claude CodeAgent at 60.4%. Below that tier, movement exists but stays contained: JunieAgent holds sixth place, Cursor's agent drops slightly in the field but remains measurable, and the full-parameter models like Qwen variants continue their lower placements. The benchmark methodology itself remains sound, it tests actual repository-level software engineering tasks with controlled evaluation conditions and reports confidence intervals that reflect genuine measurement uncertainty rather than false precision. What's notable is the absence of dramatic reshuffling despite the dense model ecosystem now tracked. The top tier separation, where Fable 5 leads by 0.7 percentage points over Grok 4.5 within their respective confidence bands, suggests these models have approached a performance ceiling on this particular task distribution, or that the benchmark is now saturating around what current agentic approaches can solve. The Artificial Analysis leaderboard, by contrast, shows far more churn across 477 entries, with Beam 1.0 entering at position 59 and every model below shifting down one rank, indicating that general instruction-following and reasoning tasks continue to stratify models more clearly than the specific coding challenge domain where the top performers cluster.

Cole Brennan